Now · May 7 – Jun 5, 2026
My Journey from full-stack engineer to AI Engineer. Apps, agents, MCP servers — and the eval harnesses, cost tracking, and fallback patterns that decide whether any of it holds up in production. This page is the live record of that month.
2 of 3 shipped
Rebuilt the site as 'AI Engineer — apps & agents' with editorial typography, a manual dark/light toggle, and this /now page as a public sprint tracker. Services paused for 30 days while the proof gets built. Foundation for everything else this month.
A small domain-specific assistant built on the Anthropic SDK and the Vercel AI SDK. Tool calling, streaming, graceful cancellation, and a real eval harness that runs in CI. The point is to show what 'production-grade' actually means — not a tutorial demo.
An open-source MCP server starter showing the 7 patterns that separate a demo from production: bearer auth with per-tool scopes, versioned tool registry, idempotency keys, structured per-call logging, deprecate-then-remove, per-tool cost instrumentation, and SDK-survivable schema validation. TypeScript on the Anthropic MCP SDK, 47 vitest cases, stdio + HTTP entry points, Claude Desktop and Cursor config snippets in /examples.
2 of 2 on the blog
A 3-axis scorecard — user value × engineering cost × LLM cost-per-request — walked through with four real candidate features. The first long-form under the AI Engineer rebrand.
The gap between an MCP demo and an MCP server you'd hand to a paying user. Auth, error handling, observability, the unglamorous parts.
2 of 4 in muscle memory
New long-form post + the build I shipped that week. No filler, no roundups.
Public services come back next month. In the meantime, intro chats are open and free.